Contradictory, My Dear Watson

Kaggle competition: Natural Language Inference (NLI) on 15 languages. Given a premise and a hypothesis, classify their relationship:

Label	Meaning
0	Entailment
1	Neutral
2	Contradiction

Dataset

• Train: 12,120 samples — 15 languages (English ~57%, + Arabic, French, Chinese, etc.)
• Test: 5,195 samples

Approach

Fine-tuning DeBERTa-v3-base (Microsoft) — strong encoder for NLI tasks. Phase 1: English-only training. Phase 2: full multilingual.

Improvements over the baseline KerasNLP notebook:

• DeBERTa-v3 instead of BERT-base-multilingual
• Stratified train/val split
• AdamW + warmup + cosine decay
• Early stopping (patience=2)
• Mixed precision (fp16)

Project structure

├── data/                  # CSVs (git-ignored)
├── notebooks/             # Exploratory notebook (KerasNLP baseline)
├── src/
│   ├── dataset.py         # PyTorch Dataset + data loading
│   ├── train.py           # HuggingFace Trainer pipeline
│   └── predict.py         # Generate submission.csv
├── configs/
│   └── config.yaml        # All hyperparameters
├── outputs/               # Checkpoints + submissions (git-ignored)
└── requirements.txt

Usage

pip install -r requirements.txt

# Train (English only)
python src/train.py --config configs/config.yaml

# Generate submission
python src/predict.py --model_dir outputs/best_model --config configs/config.yaml

Results

Model	Filter	Val Accuracy
BERT-base-multi (baseline)	All languages	~59%
DeBERTa-v3-base	English only	TBD
DeBERTa-v3-base	All languages	TBD